摘要
在数字图像中去除雨滴的干扰,对恢复图像质量有较大应用意义。随着深度学习图像去噪技术的发展,针对目前大多数去雨滴的方法恢复质量低、计算量大等问题,提出了一种基于自注意力机制的轻型图像去雨滴算法。该算法建立了一个轻量级的级联稠密残差网络(cascaded dense residual network),用于恢复被雨滴覆盖的图像信息。该网络由多个模块组成,模块间用稠密的残差和跳过连接引导图像信息的输出,以从粗略到细节的方式逐级消除图像中的雨滴,恢复雨滴下的背景信息。网络中每个模块由卷积层、非局部神经网络(non-local neural network)和递归卷积网络组成,在保证预测无雨图像的效果的同时减少参数量。实验结果表明,与AttentiveGAN等去雨滴方法相比,该算法去雨滴效果良好。该方法将自注意力机制加入级联稠密残差网络中,参数量仅为0.22 M,适用于小型嵌入式的除雨滴设备。
Removal of raindrop interference in digital images is of great significance for restoring image quality.With the development of deep learning image denoising technology,in order to solve the problems of low recovery quality and large calculation amount of most current raindrop removal methods,a light image raindrop removal algorithm based on self-attention mechanism is proposed.The algorithm establishes a lightweight cascaded dense residual network to recover image information covered by raindrops.The network is composed of multiple modules.Dense residuals and skip connections are used between the modules to guide the output of image information.The raindrops in the image are eliminated step by step from the rough to the details,and the background information under the raindrops is restored.Each module in the network is composed of convolutional layers,the non-local neural network and the recursive convolution network,which reduces the number of parameters while ensuring the effect of predicting de-raindrop images.Experimental results show that compared with the raindrop removal method such as AttentiveGAN,the algorithm has a better raindrop removal effect.In this method,the self-attention mechanism is added to the cascaded dense residual network.The parameter is the only 0.22 M,which is suitable for small embedded raindrop devices.
作者
郭嘉
蒋旻
刘双元
江佳俊
GUO Jia;JIANG Min;LIU Shuang-yuan;JIANG Jia-jun(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System(Wuhan University of Science and Technology),Wuhan 430065,China)
出处
《计算机技术与发展》
2021年第5期54-61,共8页
Computer Technology and Development
基金
国家自然科学基金项目(41571396,61702385)。
关键词
雨滴去除
深度学习
图像去噪
轻量级算法
自注意力
级联稠密残差网络
raindrop removal
deep learning
image denoising
lightweight algorithm
self-attention
cascaded dense residual network
